1. Scikit-learn: Machine learning in Python
Brian Holt

2. Project goals and vision
Machine learning for applications
Ease of use
Light and easy to install package
A general-purpose high level language: Python
High standards
State-of-the-art algorithms
High quality bindings: performance and fine control
Open Source
BSD license
Community driven

3. API and design 1 Design principles Code sample
Minimise number of object interfaces
Build abstractions for recurrant usecases
Simplicity, Simplicity, Simplicity
(no framework, no pipelines, no dataset objects)
Code sample
>>> from sklearn import svm
>>> classifier = svm.SVC()
>>> classifier.fit(X_train, y_train)
>>> y_pred = clf.predict(X_test)

4. API and design 2 All objects Classification, regression, clustering
estimator.fit(X_train, y_train)
Classification, regression, clustering
y_pred = estimator.predict(X_test)
Filters, dimension reduction, latent variables
X_new = estimator.transform(X_test)
Predictive models, density estimation
test_score = estimator.score(X_test)
One day: On-line learning
estimator.refit(X_train, y_train)

5. Main features and algorithms
Supervised learning (classification and regression)
SVM high quality libsvm bindings
Generalised linear models
Least squares, ridge regression, Orthogonal matching pursuit...
Nearest Neighbours
KDTree, BallTree
Gaussian Processes
Decision tree (CART)
Coming soon: Ensembles (boosting, bagging)

6. Main features and algorithms cont.
Unsupervised learning
Gaussian mixture models
Manifold learning
Clustering
Model selection
Cross-Validation
Grid Search

7. Demos SVM classifier on iris dataset Cross validation PCA Kmeans
Compare ML algorithms

8. Conclusion Scikit-learn Well-suited for applications
Machine learning without learning the machinery
Simplicity and readability
Optimise only the critical parts Code review
Unit testing Coding standards
Well-suited for applications
Used for large datasets
Building blocks for application-specific algorithms